Given that the hybrid QSVM underperforms Random Forest substantially in accuracy, how does the proposed framework justify practical deployment, and under what workload dynamics (e.g., retraining frequency, feature drift, or dataset scale) does the quantum component provide a net advantage over classical baselines?
That’s a very fair question, and we agree with the premise. In our current experiments, Random Forest clearly outperforms the hybrid QSVM in terms of raw accuracy, and we do not claim otherwise. Our framework is not proposing to replace strong classical baselines, but to augment them under specific operational conditions.
The motivation for the quantum component is not higher accuracy on static datasets, but its potential role under dynamic cloud workloads. In environments where workload characteristics change frequently and retraining is required often, the amortized training behavior of kernel-based models becomes relevant. Once quantum kernels are computed, QSVM fitting can be relatively fast, which may be beneficial in frequent retraining scenarios.
Additionally, we observe that the hybrid QSVM maintains strong discriminative capability, with a ROC AUC of 0.88, indicating that it captures nonlinear feature interactions effectively. This can be valuable in cases where job failures arise from complex interactions among resource and scheduling features.
Practically, our framework is modular and supports adaptive α tuning. In deployment, the quantum contribution can be increased or reduced depending on workload dynamics, feature drift, and operational constraints. As quantum hardware and kernel evaluation mature, this hybrid design allows incremental adoption rather than disruptive replacement.
